In-situ prediction of reactive ion etch endpoint using neural networks

Reactive ion etching (RIE) in radio frequency glow discharges is perhaps the most popular means of achieving the level of detail necessary to pattern small geometry features in electronics manufacturing. However, the complexity of the RIE process has prompted the use of empirical models utilizing neural networks, which offer advantages in both accuracy and robustness over statistical methods. In this paper, a neural network trained to model the correlation between DC bias and etch rate was used to predict the time required to remove a specified thickness of silicon dioxide (SiO₂) in a CHF₃ /O₂ plasma. A real-time data acquisition system that transmits process conditions from a Plasma Therm 700 series RIE system was used to monitor DC bias during etching. A back-propagation neural network was trained to predict the amount of time required to etch the remaining amount of film while in the midst of etching. Inputs to the network included elapsed time during the etch run, the desired etch depth, gas flow rates, chamber pressure, and RF power. This network exhibited a 26 s RMS error on training data, and predicted the process endpoint on a set of test etch recipes with an average error of less than two minutes for a process time of about 25 min